本文主要研究机器学习方法在用哈勃参量等数据的宇宙学参数限制和FAST（五百米口径球面射电望远镜）地外文明数据处理中的应用。针对宇宙学参数限制的问题，本文注意到以下问题：宇宙学数据的误差分布（或似 然函数）无法用解析表达式精确给出，使得经典参数限制方法不够准确或失效；当数据维数较大时，消耗计算资源较多。据此首次提出结合降噪自编码器（Denoising Autoencoder，DAE）和掩码自回归流（Masked Autoregressive Flow，MAF）两种神经网络实现数据降维和“无似然”参数限制流程，该流程可绕过似然函数表达式，更准确处理误差分布。对千 DAE 数据降维，本文提出了一种适用于DAE 的创新的损失函数；对于MAF 参数限制，本文还首次指出参数的采样分布对MAF 效果的不利影响，并据此提出了一种“顺序训练方法”来减小这一影响。在传统方法未失效的模拟哈勃参量数据上进行测试表明，本文首次提出的处理流程和顺序训练方法能够准确地给出参数的限制范围，达到与传统方法可比的能力，而且本文提议应用的 DAE 可以在数据模拟量较少的情况下使 MAF 获得更好的训练结果。本文还对于这一流程具体在哈勃参量数据和 Ia  型超新星上的应用进行了讨论，并提出了新的思路，新思路需要使用的经验结果和先验知识更少。将本文的流程应用在实测数据上，得到了与经典方法一致的参数限制结果。对于哈勃参量数据的模拟，本文也首次提出使用高斯过程改进原有方法，比较表明这一方法更好地描述了哈勃参量数据测量的不确定度随红移的变化规律。针对 FAST 地外文明观测的数据处理问题，本文讨论了地球上产生的各类射频干扰信号及其特点，并据此基于聚类机器学习方法、Hough 变换图像处理等方法，给出了一种可以有效去除大部分射频干扰信号的处理方法。

This work explores the application of machine learning to cosmological constraints and FAST Searching for Extraterrestrial Intelligence (SETI). With regard to cosmological constraints, this work focuses on these problems: the errors of cosmological data cannot be accurately modeled with analytical probability distributions, making classical constraining methods less accurate or even impossible; when the dimensionality of the data is large, the constraining process is computational ex- pensive. Considering the problems, this work proposes for the first time that two neural networks, the denoising autoencoder (DAE) and the Masked Autoregressive Flow (MAF), can be used together to reduce the dimensionality of data and perform likelihood-free cosmological constraints, bypassing the direct evaluation of the likelihood and consider the distribution of the errors more accurately. In the dimensionality reduction step, this work proposes a new loss function for the DAE; in the MAF parameter constraining step, this work points out for the first time that the distribution from with parameters are sampled may have a negative impact on the training of MAF, and proposed a sequential training process to alleviate the negative impact. Evaluated on simulated Hubble parameters, the process and the sequential method first proposed in this work give accurate parameter constraints, and achieve performances that are comparable to traditional methods. The experiments also show that the MAF gets better training results for small number of simulation when the DAE is added. This work also proposes new methods on the specific application of the proposed procedure to observational Hubble parameters and the type Ia supernovae, with less prior knowledge needed. Applying the proposed method to real observational data, the work gets constraints that are consistent to relevant works. For the simulation of Hubble parameters, this work also proposes an improvement using the Gaussian process for the first time. Compared with other methods, the proposed method with Gaussian process better models the change of uncertainties with red- shifts. With regard to FAST SETI, this work discusses the properties of the radio frequency interference (RFI), and gives a procedure to remove most of the RFI, using methods including clustering algorithm and Hough transformation.